Process Automated Builder

Scope

• Build ILCD process and source datasets from one reference flow input.
• Run, resume, stop, inspect, and publish the workflow safely.
• Diagnose failures in references/SI processing, matching, unit alignment, placeholder resolution, and post-build reviews.

Execution Baseline

1. Read references/workflow-map.md and references/operations-playbook.md.
2. Bootstrap the standalone Python environment.
3. Run the wrapper with agent-provided flow input (--flow-file, --flow-json, or --flow-stdin).
4. Inspect run artifacts and continue with --resume or --publish-only when needed.

Parallel Execution Contract

• Run-level parallel: multiple flow inputs can run concurrently, but each run must use a distinct run_id.
• In-run parallel: only fan-out inside approved stage internals; stage barriers stay fixed.
• Barrier policy:
  • 01 -> 02 -> 03 strict serial.
  • 04 may fan-out over SI files.
  • 05 -> 06 -> 07 strict serial convergence.
• Single-writer rule:
  • Never let multiple agents write the same artifacts/process_from_flow/<run_id>/cache/process_from_flow_state.json.
  • Within one run_id, only one active writer process is allowed at a time.
  • Enforcement is code-level: writers acquire process_from_flow_state.json.lock before state writes.
• Flow-search parallelism:
  • 07_main_pipeline now parallelizes only flow_search requests, then applies selector and state updates in original exchange order.
  • Tune with env LCA_FLOW_SEARCH_MAX_PARALLEL (bounded by profile concurrency).

Commands

scripts/setup-process-automated-builder.sh
source .venv/bin/activate
export TIANGONG_LCA_REMOTE_TRANSPORT="streamable_http"
export TIANGONG_LCA_REMOTE_SERVICE_NAME="TianGong_LCA_Remote"
export TIANGONG_LCA_REMOTE_URL="https://lcamcp.tiangong.earth/mcp"
export TIANGONG_LCA_REMOTE_API_KEY="<your-api-key>"
export OPENAI_API_KEY="<your-openai-api-key>"
export OPENAI_MODEL="gpt-5"

scripts/run-process-automated-builder.sh --mode workflow --flow-file /abs/path/reference-flow.json -- --operation produce
scripts/run-process-automated-builder.sh --mode langgraph --flow-file /abs/path/reference-flow.json -- --stop-after matches --operation produce
scripts/run-process-automated-builder.sh --mode langgraph -- --resume --run-id <run_id>
scripts/run-process-automated-builder.sh --mode langgraph -- --publish-only --run-id <run_id> --commit
scripts/run-process-automated-builder.sh --mode langgraph -- flow-auto-build --run-id <run_id>
scripts/run-process-automated-builder.sh --mode langgraph -- process-update --run-id <run_id>

Bundled Python Scripts

• Wrapper and setup: scripts/run-process-automated-builder.sh, scripts/setup-process-automated-builder.sh
• Main chain: scripts/origin/process_from_flow_workflow.py, scripts/origin/process_from_flow_langgraph.py
• SI and references: scripts/origin/process_from_flow_download_si.py, scripts/origin/mineru_for_process_si.py, scripts/origin/process_from_flow_reference_usability.py, scripts/origin/process_from_flow_reference_usage_tagging.py
• Maintenance: scripts/origin/process_from_flow_build_sources.py, scripts/origin/process_from_flow_placeholder_report.py
• Shared helper copied for LangGraph CLI import path: scripts/md/_workflow_common.py

Runtime Requirements

• Use bundled runtime package tiangong_lca_spec/ shipped with this skill.
• Install Python dependencies via scripts/setup-process-automated-builder.sh.
• Configure flow-search MCP from env: TIANGONG_LCA_REMOTE_TRANSPORT, TIANGONG_LCA_REMOTE_SERVICE_NAME, TIANGONG_LCA_REMOTE_URL, TIANGONG_LCA_REMOTE_API_KEY.
• Configure OpenAI from env when LLM is enabled: OPENAI_API_KEY, optional OPENAI_MODEL, optional OPENAI_BASE_URL.
• Configure KB MCP from env when literature retrieval is needed: TIANGONG_KB_REMOTE_TRANSPORT, TIANGONG_KB_REMOTE_SERVICE_NAME, TIANGONG_KB_REMOTE_URL, TIANGONG_KB_REMOTE_API_KEY.
• Configure MinerU from env when SI OCR parsing is needed: TIANGONG_MINERU_WITH_IMAGE_URL, optional TIANGONG_MINERU_WITH_IMAGE_API_KEY, optional provider/model/timeout flags, optional TIANGONG_MINERU_WITH_IMAGE_RETURN_TXT (default true).

Fast Troubleshooting

• Missing process_datasets or source_datasets: verify stop_after did not stop before dataset stages.
• Too many placeholders: run through Step 6 (resolve_placeholders) and inspect cache/placeholder_report.json.
• Unit mismatch failures: inspect Step 4b flow_search.unit_check; density conversion only applies to product/waste mass<->volume mismatches.
• Slow runs: inspect cache/workflow_timing_report.json; Step 4 matching is usually the longest stage.

Load References On Demand

• references/process-from-flow-workflow.md: complete migrated workflow spec (core flow, orchestration flow, state, outputs, publishing, stop rules).
• references/workflow-map.md: standalone skill execution map (input/output contracts and run control).
• references/operations-playbook.md: operational commands for setup, run, resume, and publish.